Punch-list item 14 (ROAD-V1-C2-LOCAL-BF16) directs work that landed 2026-07-27, on a premise the same commit refuted

[localai-bot]

Punch-list item 14 in .agents/specs/roadmap-v1-completion.md §3, and the matching
Remaining to close the row cell in §2's RI table, direct a fresh implementer to:

Port device-resident sampled-token mapping to discrete CUDA (remove the measured
main-stream wait) and rerun the exact 4B series. Small.

All three clauses are stale. The record was authored from the 2026-07-25
CLAIM-LOCAL-BF16-H32-AOT checkpoint and never reconciled after the work landed.

1. The discrete-CUDA port landed 2026-07-27

deed7c2a1 — feat(engine): ENG-ASYNC-SCHED W4 discrete device-resident sampled
tokens (opt-in), and the attribution that motivated it is wrong. It adds
runner-owned device buffers for last_sampled / prefill_len /
query_start_loc / seq_lens / input_ids, and LaunchApplyLastSampledOps to
replay our condense's structural row edits on-device in stream order — i.e.
exactly the "request-compaction-safe device token map" the checkpoint asked for.

$ git log -S'AsyncDeviceInputs' --oneline
deed7c2a1 2026-07-27 feat(engine): ENG-ASYNC-SCHED W4 discrete device-resident sampled tokens (opt-in) ...
1718bf155 2026-08-05 perf(qwen3_5): device-resident sampled tokens + async batch-1 P0 fix, mirror default ON (#31)

2. It is the production default, not an opt-in

1718bf155 (2026-08-05, row/SERVE-ASYNC-LLM) flipped it ON. In the tree today,
GPUModelRunner::async_device_mirror() (src/vllm/v1/worker/gpu/runner.cpp:2430-2455)
engages on any real CUDA GPU, and names discrete first:

- DISCRETE (separate memory, !UnifiedMemory): the mirror is REQUIRED, because the host fallback would main-stream Synchronize to read the sampled ids.

VT_ASYNC_DEVICE_MIRROR=0 is the rollback; VT_ASYNC_RUNNER has defaulted ON
since the 2026-07-17 flip. The Synchronize that item 14 asks to remove survives
only on the non-CUDA host branch of sample_tokens_async
(runner.cpp:2687-2695), where it is correct and required.

3. The "measured main-stream wait" was refuted by the same commit

deed7c2a1's own message records the finding:

THE FINDING. [...] vllm-bench drives the SYNCHRONOUS LLMEngine::step() loop,
which calls sample_tokens(), not AsyncLLM's depth-2 step_with_batch_queue. So
on the benchmarked path there is no async sampler and no sample_tokens_async
synchronize to remove, and the 2026-07-25 attribution of 497
cudaStreamSynchronize calls (20.975 s, 42.20 ms/call) to that function is wrong
for this workload.

So the lever's motivating measurement does not exist on the benchmarked path.

4. The 4B series has been rerun six times since

docs/bench-evidence/ carries 20260727, 20260728, 20260729, 20260803,
20260805, 20260807. Roadmap row 2a already records the corrected
three-repetition cross-engine result (throughput 6831.71 vs 6643.40 tok/s,
1.0283x PASS) against the 2026-07-25 numbers the stale cell still quotes
(5769.99/5660.70/5849.80, 0.9864x).

The row itself is current; only these records drifted

Roadmap row 2a's next-actions and the row's live campaign spec
(specs/sm120-qwen35-pareto-2026-08-09.md, whose Ordered campaign never
mentions sampled tokens at all) both already describe the real remaining work:
same-tool interval profiling to split the residual TTFT into intake vs prefill,
then the largest attributable interval; TPOT/ITL and peak VRAM stay open.

Stale surfaces to reconcile:

• .agents/specs/roadmap-v1-completion.md:39 (RI table) and :206 (punch-list item 14)
• .agents/benchmark-record.md — Host-weight ownership key, next-actions cell
• .agents/coordination.md — the 2026-07-25 CLAIM-LOCAL-BF16-H32-AOT bullet's
  "Next owner must..." directive

This is a records-only fix: no src/, include/ or tests/ change, and no row
lifecycle change (ROAD-V1-C2-LOCAL-BF16 stays GATING).